This script illustrates the use of an existing set of prototypes to compute image features. This data is then used to evaluate classifier performance.

evaluate-classifier-for-protos.py

from glimpse.experiment import *

def make_model_for_protos(protos):
    """Create a model for an existing set of prototypes."""
    model = Model()
    # Here, we assume that all prototypes have the same spatial extent.
    model.s2_kernels = [protos]
    return model

def train_classifier(model, pool, images, labels):
    """Train a classifier on features extracted from a set of images.

    Features are extracted using an existing model.

    """
    layer = model.LayerClass.C2
    # Compute model activity for each image.
    images = map(model.MakeState, images)
    builder = Callback(BuildLayer, model, layer, save_all=False)
    states = pool.map(builder, images)
    # Create feature vectors from model activity.
    features = ExtractFeatures(layer, states)
    # Train the classifier.
    clf = FitClassifier(features, labels)
    return clf

def test_classifier(model, pool, clf, images, labels):
    """Test an existing classifier on features extracted from a set of images.

    Features are extracted using an existing model.

    """
    layer = model.LayerClass.C2
    # Compute model activity for each image.
    images = map(model.MakeState, images)
    builder = Callback(BuildLayer, model, layer, save_all=False)
    states = pool.map(builder, images)
    # Create feature vectors from model activity.
    features = ExtractFeatures(layer, states)
    # Test the classifier.
    acc,_ = ScoreClassifier(features, labels, clf=clf)
    return acc

def example_protos():
    """Create an arbitrary set of example prototypes."""
    from glimpse import corpora
    corpus = os.path.dirname(corpora.__file__) + "/data/easy"
    exp = ExperimentData()
    SetModel(exp)
    SetCorpus(exp, corpus)
    MakePrototypes(exp, 10, 'imprint')
    # Here, we assume that all prototypes have the same spatial extent.
    protos = exp.extractor.model.s2_kernels[0]
    return protos

def example_corpus():
    """Return information about an arbitrary image corpus."""
    from glimpse import corpora
    corpus = os.path.dirname(corpora.__file__) + "/data/easy"
    exp = ExperimentData()
    SetCorpus(exp, corpus)
    return exp.corpus.paths,exp.corpus.labels

def main():
    """Illustrate the use of the above functions."""
    # Generate a bunch of example data.
    images,labels = example_corpus()
    mask = ChooseTrainingSet(labels, train_size=0.5)
    train_images = images[mask]
    test_images = images[~mask]
    train_labels = labels[mask]
    test_labels = labels[~mask]
    protos = example_protos()
    # Run the functions on the example data.
    model = make_model_for_protos(protos)
    pool = MakePool()
    clf = train_classifier(model, pool, train_images, train_labels)
    acc = test_classifier(model, pool, clf, test_images, test_labels)
    print "learned classifier with accuracy: %s" % acc

if __name__ == "__main__":
    main()